VARIABLE PRIORITIZATION IN NONLINEAR BLACK BOX METHODS: A GENETIC ASSOCIATION CASE STUDY¹

Area of Science:

• Statistical genetics
• Bioinformatics
• Machine learning

Background:

• Nonlinear and nonparametric regression models are crucial for analyzing complex biological data, especially in statistical genetics.
• Identifying important predictor variables, particularly genetic variants with nonlinear interactions, remains a challenge.
• Existing methods may not fully capture the importance of variables involved in complex, covarying relationships.

Purpose of the Study:

• To introduce a novel and interpretable measure, RelATive cEntrality (RATE), for variable selection in nonlinear regression.
• To prioritize genetic variants based on both marginal importance and significant covarying relationships.
• To provide a method that explains the improved predictive accuracy of nonlinear models in complex genetic architectures.

Main Methods:

• Development of the RelATive cEntrality (RATE) measure.
• Application of RATE within Bayesian Gaussian process regression models.
• Illustration using simulations and two real genetic association mapping studies.

Main Results:

• The RATE measure effectively prioritizes candidate genetic variants by considering their marginal and covarying importance.
• Bayesian Gaussian process regression combined with RATE demonstrates improved performance in identifying relevant genetic variants.
• Simulations and real data analyses confirm the utility of RATE in explaining the enhanced predictive power of nonlinear models.

Conclusions:

• The RATE measure offers a robust and interpretable approach to variable selection in nonlinear and nonparametric regression.
• RATE facilitates a deeper understanding of complex genetic architectures by highlighting important covarying relationships among variants.
• This methodology has significant implications for statistical genetics and the analysis of complex phenotypes.